Reproduce results on MMStar Benchmark Datasets.

[1359347500cwc]

I tried to use the code for llama.vision and the weights you provided to evaluate on the MMStar benchmark dataset in VLMEvalKit , I found that the results are inconsistent. The results I obtained are as follows:

Overall: 0.4806666666666667

Coarse perception: 0.56

Fine-grained perception: 0.448

Instance reasoning: 0.52

Logical reasoning: 0.492

Math: 0.544

Science & technology: 0.32

This shows a significant gap compared to the official result of 59.53.